figured out how to precisely add the PWM with its delayed version and integrate them nicely together:

should try adding a DC-blocking capacitor before input to opamp

Nothe: the RC lowpasses are unnecessary for reconstruction, or their cutoff freq could even be made lower closer to 20kHz if desired for smoothness. The integrator is the one crucial part.

And for feedback, the RC cutoff is the main thing that influences the oscillation frequency (other than other circuit dealys). Thus, a Potentiometer could be used for the resistor to tweak the oscillation frequency to be exactly 96kHz.

tweaks

maybe there is some other way to encode 2nd derivative other than just using 1st derivative as input or mix it in with some CR network when combining with feedback result.

Above was output of adding RC-lowpassed both Q and delayed Q summed and integrated.

Below is just Q's output:

Below is the integrator's output:

sim with tighter integral output compared to input

Q out:

integrated out:

doesn't seem better.